Upstream 2026.06.15

[handstandsam]

Summary

Syncs the latest internal changes into the open-source repo (through 2026.06.15).
52 files change (9 new). The headline themes for anyone consuming Trailblaze:

• assertVisible can now check an element's text, not just its presence. Pass
  the new optional expectedText to assert the rendered value (e.g. "expect the
  checkout button to read Charge $5.00", "expect status to be Active"). Volatile
  trailing state like a live item count (Review sale\n3 items) is detected at
  capture time and matched loosely so it doesn't make replays brittle.
• New clearText tool wipes the focused text field completely before you type.
  The bare Maestro eraseText silently caps at 50 characters on every driver, so
  long values (passwords, addresses, search queries) used to survive a "clear" and
  concatenate onto the next input. clearText reads the field's real length and
  erases exactly that many characters.
• The agent stops tapping the wrong thing when its target is off-screen. When a
  step names a target that isn't among the tappable elements, the agent now scrolls
  to reveal it (if there's a scroll affordance) or stops with a clear "you may be on
  the wrong screen" signal — instead of tapping an unrelated nearby element and
  marching on.
• Soft-keyboard mis-taps on Android are caught at the point of impact. If a
  resolved tap coordinate falls inside the on-screen keyboard's bounds (which happens
  on modal screens that swallow the Back gesture and refuse to dismiss the keyboard),
  the tap now fails loudly with a clear message instead of silently landing on the
  keyboard.
• Multiple Trailblaze daemons can run on one host without stepping on each other.
  A new TRAILBLAZE_HOME environment variable isolates each daemon's state directory
  (logs, TLS keystore) so concurrent daemons don't race a shared keystore on boot.

What changed, by area

Agent behavior & verification

• Target discipline (BlazeGoalPlanner, ReflectionNode): a new pre-execution
  check (detectTargetMissingRecovery) keeps the agent from tapping a distractor
  element when the step's named target isn't actually on screen — it scrolls toward a
  revealed affordance or surfaces a wrong-screen signal.
• Verify steps must leave a real assertion (StepToolSet): the verify prompt now
  requires the agent to call an assertion tool that returns success (not just "look
  at the screen"), so the captured trail has a concrete assertion to replay
  deterministically.

assertVisible text verification + clearText

• AssertVisibleTrailblazeTool / AssertVisibleBySelectorTrailblazeTool gain an
  optional expectedText plus a TextMatchMode (EXACT / PREFIX / REGEX).
  Default stays EXACT, and trails recorded before this change deserialize unchanged.
• VolatileTextDetector rewrites a captured expectedText so a trailing volatile
  item-count subtitle is tolerated at replay (changed or gone), while the stable
  text stays an exact requirement.
• New ClearTextTrailblazeTool (registered in the core_interaction toolset) and its
  generated reference doc.

Android accessibility driver

• IME-occlusion safety net (AccessibilityDeviceManager,
  TrailblazeAccessibilityService): pre-tap check against the live keyboard window
  bounds (with a dumpsys fallback when window enumeration is degraded); hideKeyboard
  is now best-effort on its post-check, with an env-gated SHOW_MODE_HIDDEN dismissal
  path (TRAILBLAZE_IME_DISMISS_VIA_SHOW_MODE) for Compose modals that consume Back.
• Compact element list (AndroidCompactElementList): emits a non-tappable
  "(scroll … to reveal)" affordance for off-screen editable fields, and recurses into
  Compose interop containers that are marked invisible but whose children (modal
  content, overlay buttons) are individually visible.
• Selector stability (TrailblazeNodeSelectorMinimizer, generator): an
  index-carrying selector keeps its most-stable anchor instead of collapsing to a
  naked positional index (which shifts whenever anything before the target changes);
  a genuinely attribute-less node still falls back to a bare index.
• Tap-divergence logging (TapTrailblazeTool): logs when the OS hit-test winner
  at a tapped coordinate differs from the element the agent picked, making a class of
  "looked-correct" selector mismatches observable in every capture.

Concurrency & daemon isolation

• TRAILBLAZE_HOME honored by getDefaultAppDataDirectory, SslConfig, and the MCP
  stdio proxy so concurrent daemons isolate their state dirs.
• DaemonScriptedToolBundler: the stale-wrapper sweep only deletes wrapper files
  older than the esbuild timeout, so daemons in separate JVMs sharing a source dir
  don't delete each other's in-flight wrappers.
• RunYamlRequestHandler: an originating HTTP-call cancellation now cancels the
  launched job rather than stalling to the handler timeout cap.

Security & install hygiene

• AndroidHostAdbUtils.redactSecretsForLog: LLM provider auth tokens passed as
  trailblaze.llm.auth.token.<provider> args are redacted from logged adb shell
  commands (CI archives those logs as downloadable artifacts).
• PrecompiledApkInstaller: the on-device APK-SHA marker write now uses printf
  instead of a sh -c "echo … > file" wrapper that got mangled by the argv space-join
  — the marker was silently blank, defeating server reuse and forcing a reinstall on
  every run. Now written verbatim and read back to verify.

Scripted-tool (QuickJS) execution

• QuickJsToolSerializer / QuickJsTrailblazeTool / LazyYamlScriptedToolRegistration:
  thread the session binding into the decoded tool so a nested client.callTool(...)
  from inside a bundled handler resolves its execution context on the LLM dispatch
  path (previously returned "no execution context installed").

Desktop run-config & reporting

• TrailblazeDeviceManager: per-run overrides from the Run-config dialog (self-heal,
  recorded-steps replay, max LLM calls, memory seeds, video/logcat/iOS-log/network
  capture). All default to prior behavior, so existing callers are unaffected.
• MainTrailblazeApp: an optional extraDaemonRoutes hook lets a downstream desktop
  edition register extra daemon endpoints without the lower layer depending on them.
• TrailblazeServerState.captureAnalytics, LogsRepo (retry surfacing a newly
  detected session until its first log is written; reuse cached info for completed
  sessions), and test-result priority (P0/P1/P2) + Android/iOS build-version fields.

Dependencies, examples, docs

• Add micrometer 1.15.12 and ktor-client-mock; refresh the runtime-classpath
  baselines (micrometer 1.13.4 → 1.15.12).
• examples/playwright-electron: a provision-electron.sh step fetches the Electron
  platform binary explicitly (npm ≥ 11.16 no longer runs Electron's postinstall).
• Regenerated tool-reference docs (assertVisible expectedText, new clearText,
  updated toolset counts).

Test plan

• OSS CI green (build + unit tests across modules)
• spotlessApply clean
• Sensitive-terms scan clean (run pre-sync)